SaltiNet: Scan-path Prediction on 360 Degree Images using Saliency\n Volumes

Abstract

We introduce SaltiNet, a deep neural network for scanpath prediction trained\non 360-degree images. The model is based on a temporal-aware novel\nrepresentation of saliency information named the saliency volume. The first\npart of the network consists of a model trained to generate saliency volumes,\nwhose parameters are fit by back-propagation computed from a binary cross\nentropy (BCE) loss over downsampled versions of the saliency volumes. Sampling\nstrategies over these volumes are used to generate scanpaths over the\n360-degree images. Our experiments show the advantages of using saliency\nvolumes, and how they can be used for related tasks. Our source code and\ntrained models available at\nhttps://github.com/massens/saliency-360salient-2017.\n